Open Access
Table 3.
CNN architectures.
| Layer type | # Param. | Output shape | Properties |
|---|---|---|---|
| Input | 0 | (1,N,N) | |
| Convolutional | 1952 | (16,N/2,N/2) | 2 pixels stride, |
| 16 filters (11,11) | same padding, | ||
| Leaky ReLU act. | |||
| Batch Norm. | 2N | (16,N/2,N/2) | |
| Dropout | 0 | (16,N/2,N/2) | 50% |
| Convolutional | 12 832 | (32,N/4,N/4) | 2 pixels stride, |
| 32 filters (5,5) | same padding, | ||
| Leaky ReLU act. | |||
| Batch Norm. | N | (32,N/4,N/4) | |
| Dropout | 0 | (32,N/4,N/4) | 50% |
| Convolutional | 18 496 | (64,N/8,N/8) | 2 pixels stride, |
| 64 filters (3,3) | same padding, | ||
| Leaky ReLU act. | |||
| Batch Norm. | N/2 | (64,N/8,N/8) | |
| Dropout | 0 | (64,N/8,N/8) | 50% |
| Flatten | 0 | (N2) | |
| Dense | (N2+1)⋅64 | (64) | 64 units, |
| ReLU act. | |||
| Dropout | 0 | (64) | 30% |
| Dense | 2080 | (32) | 32 units |
| ReLU act. | |||
| Dropout | 0 | (32) | 30% |
| Dense | 33 | (1) | 1 unit, |
| sigmoid act. | |||
Notes. The columns are the name of the Keras layer (and the filters for the convolutional layers), the number of trainable parameters, output, and hyper-parameters for each layer. N represents the size of the input images which is 120, 96, 80, and 64 for z-bins 1, 2, 3, and 4, respectively.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.